Jargon is the biggest barrier to entry. Here are the terms that unlock 90% of AI articles and docs — one honest line each.
Core ML
- Model — the learned function that maps inputs to outputs.
- Weights / parameters — the numbers inside the model that get tuned during training. "70B parameters" = 70 billion of them.
- Feature — one input variable (age, pixel, word).
- Label / target — the correct answer you train against.
- Training / inference — learning the weights vs. using the trained model to predict.
- Epoch — one full pass through the training data.
- Batch — a chunk of examples processed together.
- Learning rate — how big a step each update takes.
- Loss — a number measuring how wrong the model is.
- Overfitting — memorising the training data instead of learning the pattern (great on train, bad on new data).
- Generalisation — performing well on data it has never seen. The actual goal.
Deep learning
- Neural network — layers of simple units (neurons) that transform data.
- Activation function — adds non-linearity so networks can learn complex shapes (ReLU, sigmoid).
- Backpropagation — the algorithm that computes gradients through all layers.
- GPU / TPU — hardware that does the massive matrix math fast.
- CNN / RNN / Transformer — architectures for images / sequences / (now) everything.
LLM / GenAI
- Token — a chunk of text (~¾ of a word) the model reads and predicts.
- Embedding — text/data turned into a vector of numbers that captures meaning.
- Prompt — the input you give an LLM.
- Context window — how many tokens the model can "see" at once.
- Hallucination — the model stating something false but confident.
- Fine-tuning — further-training a base model on your specific data.
- RAG — retrieving relevant documents and feeding them to the LLM for grounded answers.
- Temperature — randomness of the output; 0 = deterministic, high = creative.
Bookmark this — you will return to it. Deep dives on each in the tracks below.